Automatic Detection and Characterization of Obstructive Sleep Apnea Using Computer Vision

Umaer Rashid Hanif

Research output: Book/ReportPh.D. thesis

231 Downloads (Pure)

Abstract

Background: Obstructive sleep apnea (OSA) is characterized by recurrent upper airway collapse during sleep and affects up to one billion people worldwide. OSA is associated with increased risk of cardiovascular diseases, stroke, and all-cause mortality. Diagnosis and treatment of OSA are crucial for long term health and a reduced economic burden. However, the gold-standard polysomnography (PSG) is time-consuming, expensive, and requires a great amount of manual labor.
Objective: The objective of this PhD project was to invent fast, cheap, and data-driven screening and scoring systems for OSA based on imaging data. Imaging data can be captured much faster than overnight data collection for detection and characterization of OSA using computer vision. 
Methods: Two systems were proposed: 1) an automatic screening system which utilizes 3D craniofacial scans to estimate apnea-hypopnea index (AHI), which measures OSA severity, and 2) an automatic scoring system which utilizes drug-induced sleep endoscopy (DISE) videos to estimate sites of upper airway collapse and obstruction degrees in OSA patients. The main components in both systems were convolutional neural networks, which were trained and evaluated using two different datasets consisting of 1) 1366 3D craniofacial scans collected across 11 different sleep clinics, and 2) 281 DISE videos collected across two sleep clinics and scored by three different surgeons.
Results: For AHI estimation based on 3D craniofacial scans, a mean absolute error of 11.38 events/hour and a Pearson correlation of 0.4 were obtained. Subjects were classified as normal or with OSA with an accuracy of 67%, which was higher than using screening questionnaires. The model’s performance was comparable to three sleep specialists, and its performance increased further by adding demographics and questionnaires as features. For automatic scoring of DISE in OSA, a mean F1 score of 70% was obtained across four upper airway sites (velum: 85%, oropharynx: 72%, tongue base: 57%, epiglottis: 65%) with respect to obstruction degrees (0, 1, or 2). 
Conclusion: The proposed automatic screening system for detection of OSA based on 3D craniofacial scans has the potential to fulfill the need for a fast and cheap screening method for OSA in clinical practice. The proposed automatic scoring system for DISE videos has the potential to provide consistent scoring without bias, which can aid surgeons in interpretation of DISE and result in improved treatment outcomes for OSA patients.
Original languageEnglish
PublisherDTU Health Technology
Number of pages200
Publication statusPublished - 2022

Fingerprint

Dive into the research topics of 'Automatic Detection and Characterization of Obstructive Sleep Apnea Using Computer Vision'. Together they form a unique fingerprint.
  • Predicting Sleep Disordered Breathing from 3D Craniofacial Imaging

    Hanif, U. R. (PhD Student), Karstoft, H. (Examiner), Penzel, T. (Examiner), Sørensen, H. B. D. (Main Supervisor), Mignot, E. (Supervisor) & Jennum, P. J. (Supervisor)

    01/03/201916/01/2023

    Project: PhD

Cite this